Unified Sensitivity Suites and Evidence Inventories

Orchestration without automatic robustness claims

gp3bayes already provides prior sensitivity, power scaling, PSIS-LOO, structural sensitivity, group-deletion sensitivity, coding/scaling variants, duration-unit invariance and exact K-fold validation. Version 0.2.0 adds a thin orchestration layer so these results can be planned and collected without turning them into an automatic “robust/not robust” verdict.

Declare a suite before running it

plan <- create_sensitivity_suite_plan(
  prior_scale = TRUE,
  powerscale = TRUE,
  psis_loo = TRUE
)
plan
#> <gp3bayes_sensitivity_plan>
#>   Prior-scale refit: TRUE
#>   Power-scaling: TRUE
#>   PSIS-LOO: TRUE
#>   Random-slope plan: FALSE
#>   Group-deletion plan: FALSE

Creating the plan runs nothing. Expensive components only run when run_sensitivity_suite() receives both a fitted model and an explicit plan.

suite <- run_sensitivity_suite(
  fit,
  plan,
  stop_on_error = FALSE
)

summarise_sensitivity_suite(suite)
plot(suite)

Structural sensitivity can be declared using the package’s existing governed plans:

random_slope_plan <- create_random_slope_sensitivity_plan(specification)
group_plan <- create_group_deletion_sensitivity_plan(
  specification,
  group = "participant",
  units = c("p001", "p002")
)

plan <- create_sensitivity_suite_plan(
  prior_scale = TRUE,
  psis_loo = TRUE,
  random_slope_plan = random_slope_plan,
  group_deletion_plan = group_plan
)

Evidence is an inventory

Already-computed results can be collected into one review object.

evidence <- collect_model_evidence(
  fit = fit,
  design = design,
  diagnostics = diagnostics,
  posterior = posterior,
  ppc = ppc,
  estimands = estimands,
  loo = loo_result,
  sensitivity = suite,
  manifest = frozen_manifest
)

evidence
plot(evidence)

Reports require an explicit file path:

report <- tempfile(fileext = ".md")
create_model_evidence_report(evidence, report)
unlink(report)

The inventory deliberately withholds aggregate adequacy, robustness, causal, and model-selection claims. Different evidence components answer different questions and can disagree without being collapsed into a single score.